Thesis REDES NEURONALES DE CONVOLUCIÓN APLICADAS A LA SEGMENTACIÓN DE LESIONES DE ESCLEROSIS MÚLTIPLE EN IMÁGENES DE RESONANCIA MAGNÉTICA 3D
Loading...
Date
2017
Authors
NARANJO COLLAO, RODRIGO ESTEBAN
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La Esclerosis Múltiple es una enfermedad crónica que afecta al sistema nervioso central. Eldiagnóstico de esta enfermedad se basa en el uso de Imágenes de Resonancia Magnética para ladetección las lesiones propias de esta. Este proceso de detección suele ser realizado de maneramanual por expertos, sin embargo, existe el interés en automatizar esta tarea con el fin de agilizar elproceso de diagnóstico y seguimiento de esta enfermedad. Con este fin, se han propuesto múltiplestécnicas automaticas de segmentación de imágenes que permitan detectar de manera efectiva estaslesiones. En esta memoria, se estudiará el desempeño de dos técnicas del Estado del Arte aplicadasal problema de detección de lesiones de Esclerosis Múltiple en Imágenes de Resonancia Magnética.La primera técnica se basa construcción de un Modelo de Mezcla de Gaussianas, con el fin dedetectar outliers en cada imagen. Luego, se aplica una serie de reglas basadas en el conocimientoexperto de las lesiones de Esclerosis Múltiple para descartar aquellos outliers del modelo que nopresenten características propias de las lesiones. Esta técnica permite obtener una segmentación delas lesiones de manera muy eficiente en términos de tiempo, sin embargo sufre del problema desubestimar de gran manera el tamaño de las lesiones. Además de esto, las lesiones segmentadassuelen corresponder a regiones de materia gris del cerebro. En algunos casos es posible detectarestos falsos positivos, sin embargo el conjunto de reglas aplicado no logra eliminar todos estoserrores sin comprometer la calidad de la segmentación.La segunda técnica consiste en el entrenamiento de una Red Neuronal Convolucional que logresegmentar las lesiones en base a un conjunto de imágenes de entrenamiento. Esta técnica es capazde aprender una serie de filtros que, aplicados a pequeñas secciones de la imagen denominadas"parches", permite obtener una serie de características que permite clasificar cada vóxel de la imagencomo lesión o tejido saludable. Este tipo de redes requiere de una gran cantidad de tiempo paraprocesar un gran número de imágenes, por lo cual se aplicaron técnicas de submuestreo de parchespara disminuir el tiempo de entrenamiento. Los resultados obtenidos muestran que las redes soncapaces de obtener resultados similares en calidad a los algoritmos del Estado del Arte en la base dedatos trabajada, sin embargo su capacidad predictiva está limitada por el muestreo inicial de parches.
Multiple Sclerosis is a chronic disease which aects the central nervous system. The diagnosis ofthis disease is based on the use of Magnetic Resonance Imaging for the detection of lesions causedby it. The process of detection is usually performed manually by experts, however, there existsinterest in the automation of this task in order to speed up the process of diagnosis and monitoringof this disease. To this end, multiple automated segmentation techniques have been proposed toeectively detect these lesions. In this undergraduate thesis, the performance of two State of the Arttechniques applied to the problem of Multiple Sclerosis lesion detection in Magnetic ResonanceImaging will be studied.The first technique is based on the construction of a Gaussian Mixture Model, with the goal ofdetecting outliers on each image. Then, a series of rules based on expert knowledge of MultipleSclerosis lesions is applied to discard the outliers which don’t present characteristics said lesions.This technique allows to obtain a segmentation of the lesions in a very ecient way in regards oftime, however it suers from the problem of greatly understimating the size of the lesions. In additionto this, segmented lesions usually correspond to gray matter regions of the brain. In some cases it’spossible to detect these false positives, however the applied ruleset is incapable of eliminating allthese errors without compromising the quality of the segmentation.The second technique consists in training a Convolutional Neural Network which learns tosegment the MS lesions from an image training set. This technique is capable of automaticallylearning a series of filters which, when applied to small sections of the image called "patches",allows to obtain a series of characteristics of the image which can be used to classify each voxelof the image as lesion or healthy tissue. This type of networks requires a large amount of time toprocess large amounts of training patches, which is why a subsampling technique was applied todecrease training time. The results obtained show that CNNs are capable of obtaining performancesimilar to that of State of Art algorithms on the dataset utilized, however its predictive capabilitiesare limited by the patch subsampling method.
Multiple Sclerosis is a chronic disease which aects the central nervous system. The diagnosis ofthis disease is based on the use of Magnetic Resonance Imaging for the detection of lesions causedby it. The process of detection is usually performed manually by experts, however, there existsinterest in the automation of this task in order to speed up the process of diagnosis and monitoringof this disease. To this end, multiple automated segmentation techniques have been proposed toeectively detect these lesions. In this undergraduate thesis, the performance of two State of the Arttechniques applied to the problem of Multiple Sclerosis lesion detection in Magnetic ResonanceImaging will be studied.The first technique is based on the construction of a Gaussian Mixture Model, with the goal ofdetecting outliers on each image. Then, a series of rules based on expert knowledge of MultipleSclerosis lesions is applied to discard the outliers which don’t present characteristics said lesions.This technique allows to obtain a segmentation of the lesions in a very ecient way in regards oftime, however it suers from the problem of greatly understimating the size of the lesions. In additionto this, segmented lesions usually correspond to gray matter regions of the brain. In some cases it’spossible to detect these false positives, however the applied ruleset is incapable of eliminating allthese errors without compromising the quality of the segmentation.The second technique consists in training a Convolutional Neural Network which learns tosegment the MS lesions from an image training set. This technique is capable of automaticallylearning a series of filters which, when applied to small sections of the image called "patches",allows to obtain a series of characteristics of the image which can be used to classify each voxelof the image as lesion or healthy tissue. This type of networks requires a large amount of time toprocess large amounts of training patches, which is why a subsampling technique was applied todecrease training time. The results obtained show that CNNs are capable of obtaining performancesimilar to that of State of Art algorithms on the dataset utilized, however its predictive capabilitiesare limited by the patch subsampling method.
Description
Catalogado desde la version PDF de la tesis.
Keywords
ESCLEROSIS MULTIPLE , IMAGEN DE RESONANCIA MAGNETICA , MODELO DE MEZCLA DE GAUSSIANAS , REDES NEURONALES CONVOLUCIONALES